AI News · 3 min read
Baidu ERNIE-5.5-Flash-Base: 280B MoE weights drop
Baidu released ERNIE-5.5-Flash-Base on Nov 5, 2026 — 280B/14B MoE, 33.5T tokens, 128K train / 1M fine-tune. Open weights under a custom license, not Apache.
What Baidu released on November 5, 2026
Baidu published ERNIE 5.5 Flash Base ↗ on November 5, 2026, with weights on Hugging Face at baidu/ERNIE-5.5-Flash-Base ↗ (FP8) and baidu/ERNIE-5.5-Flash-Base-BF16 ↗.
Card highlights (vendor numbers, not our lab):
280B total parameters / 14B activated per token (sparse MoE)
Pre-trained from scratch on 33.5T tokens
Training context up to 128K; fine-tuning and inference support 1M context
Checkpoint is end of pre-training — no mid-training or post-training yet
FP8 + BF16 safetensors; served via completions APIs in SGLang / vLLM (base model, not a chat template product)
Baidu says a post-trained ERNIE 5.5 Flash is coming soon
Architecture notes Baidu emphasizes: hybrid MLA + Heavily Compressed Attention, Multi-Router Latent MoE, and efficiency co-designed with the training stack. Treat those as the vendor’s design claims.
License: open weights ≠ open source
Weights ship under the ERNIE 5.5 Flash Base Model License Agreement v1.0 (Nov 5, 2026) — not Apache-2.0. Read the full LICENSE ↗ before commercial use. Material points from that agreement:
Commercial use is allowed below a US$50 million “Annual Revenue of Related Products or Services” threshold; at or above that threshold you need separate authorization from Baidu.
Territory in the agreement excludes the EU, UK, and South Korea for authorized use/distribution under this grant.
Acceptable-use and derivative-model rules apply; do not treat this like a permissive OSS drop.
So: downloadable weights, not a classic open-source license. Same category of “open weights with commercial tripwires” as other 2026 MoE drops — compare the shape of the deal to Kimi K3.1 Pro’s weight release (different vendor, different revenue / MaaS hooks), without inventing a winner.
What people are saying on X
Official Baidu ERNIE account:
“Introducing ERNIE-5.5-Flash-Base, the pre-trained base of ERNIE 5.5 Flash before any mid- or post-training. The post-trained ERNIE 5.5 Flash is coming soon. • 280B total / 14B activated MoE • 33.5T pre-training tokens • 128K training context, 1M for fine-tuning and inference”
— @Baidu_ERNIE ↗, Nov 5, 2026 (~123 likes at fetch)
We did not find a strong Reddit megathread specific to this base drop at research time, so we are not inventing Reddit quotes. Official blog + HF card + the X post above are the desk’s sources. Press write-ups (including MarkTechPost-class coverage) track the same vendor numbers; prefer the Baidu / HF primary sources for specs and license text.
Our take: who should care
Researchers and fine-tuners: this is a base checkpoint. Expect completions-style behavior, not a polished chat assistant, until you (or Baidu’s upcoming post-trained Flash) add instruction / RL stages. Baidu’s own quick-start uses completions, not chat.
Open-weights shoppers comparing MoEs: Baidu reports its own unified eval table vs DeepSeek-V4-Flash-Base, Hy3-preview-Base, and MiMo-V2.5-Base — those are Baidu’s numbers under Baidu’s harness. We are not reprinting a scoreboard as independent truth. Run your own domain suite after you clear license + hardware.
Hardware reality: 280B-class MoE weights are a cluster download, not a laptop afternoon. Factor FP8 vs BF16 disk, expert-parallel serving, and the 1M context claim as an inference/fine-tune ceiling, not a free lunch on one GPU.
License counsel first: if related product revenue approaches $50M, or you operate in the excluded Territory, stop and get authorization / local counsel — do not assume “Hugging Face download = free commercial forever.”
For the wider open-model landscape, see best open-source LLMs 2026, our DeepSeek review 2026, and Llama 4 review 2026. Adjacent weight-drop coverage: Kimi K3.1 Pro open weights.
AdSense-safe note: we are documenting Baidu’s release and license shape — not selling hosting, fine-tunes, or any Baidu plan.